224: AI and Computational Pathology in Breast Cancer Care

Paper Discussed in this Episode: How artificial intelligence applied to digital pathology could guide treatment personalization in breast cancer. T. Ruelle, T. Grinda, L. Del Mastro, M. Lacroix-Triki, B. Pistilli & G. Gessain. ESMO Real World Data and Digital Oncology 2026.

Episode Summary: In this journal club episode, we step into the reality of computational pathology and explore how artificial intelligence is fundamentally transforming breast cancer diagnostics. We examine a comprehensive review detailing how AI not only assists overburdened healthcare systems but also unlocks invisible genomic data straight from a standard $5 hematoxylin-eosin (H&E) glass slide. What happens when a machine can predict complex DNA mutations just by evaluating the structural architecture of cells?

In This Episode, We Cover:

• The Diagnostic Bottleneck: Understanding the critical worldwide shortage of pathologists colliding with a projected 3.2 million global breast cancer diagnoses by 2050, and why the system is under unprecedented strain.

• The Biomarker Battle: Why the human visual cortex struggles to quantify faint immunohistochemistry stains, and how AI acts as a perfect "digital colorimeter". We discuss its near-perfect concordance in assessing crucial biomarkers like Ki-67, ER, PR, PD-L1, and the newly established HER2-low status.

• Seeing the Invisible (Predictive AI): How deep learning transcends visual diagnostics to predict treatment outcomes, such as a patient's response to neoadjuvant chemotherapy. We also discuss AI's ability to infer Homologous Recombination Deficiency (HRD) and BRCA1/2 mutations by identifying macroscopic footprints like laminated fibrosis.

• Decoding Genomic Assays: The potential to replace expensive, tissue-consuming genomic tests like Oncotype DX with AI models (such as Orpheus) that predict recurrence risk straight from digitized slides, achieving accuracy that rivals the tests themselves.

• Roadblocks to Reality: The major clinical friction preventing global rollout. We discuss the steep infrastructure costs of whole-slide scanners, the danger of AI bias across diverse hospital datasets, and the ethical "black box" problem requiring the evolution of transparent, agent-based AI.

Key Takeaway: Computational pathology is moving far beyond basic diagnostic assistance. By successfully reading the structural language of biology, AI proves it can extract costly, invisible molecular data from standard biopsies, fundamentally changing the economics and accessibility of global personalized healthcare

Get the "Digital Pathology 101" FREE E-book and join us!

Transcript

You know, for over a century, the absolute cornerstone of diagnosing cancer has been um something surprisingly low tech.

Yeah. Literally just a piece of human tissue on a glass slide,

right? Sliced incredibly thin, stained with two basic dyes, hemattoxylin and eioin, giving it that classic pink and purple look. I mean, it's a cheap, reliable window into the human body.

It really is. But there is a massive shift happening right now with that simple technology.

Exactly. That basic $5 glass a slide is hiding secrets that the human eye simply cannot process

and it forces a complete rethinking of what a biopsy actually is. We're entering a paradigm where that slide is no longer just a static image.

Right. It's more like a dense landscape. Yeah. You know, full of vast amounts of invisible predictive data.

Exactly. Just waiting for the right tool to decode it.

Well, welcome trailblazers to the digital pathology podcast. Today we are doing a special journal club deep dive into the source material

and we've got a really fascinating paper to explore today.

We do. It's titled how artificial intelligence applied to digital pathology could guide treatment personalization in breast cancer. It was authored by T Ruel and colleagues and published in the journal ESMO real world data in digital oncology um back in 2025.

It's a great piece of research.

Right. So our mission today is to explore the frontier of computational pathology. We want to see how the marriage of digitized slides and artificial intelligence is stepping up to rescue a really overwhelmed health care system

and we really need to emphasize how much strain that system is under right now.

Yeah, let's talk about that. Why is this AI shift so desperately needed?

Well, to understand the tech, you first have to look at the structural breaking point in modern medicine. Back in 2022, breast cancer diagnoses hit 2.3 million globally.

Wow. And it's just going up, right?

Exactly. By 2050, demographic models project that number will reach 3.2 million. But, you know, the true crisis isn't solely the rising number of patients,

right? It's the doctors, too.

Yeah, there's a worldwide critical shortage of pathologists available to actually read those biopsies. We're facing a severe diagnostic bottleneck.

So, the math just doesn't add up. We have more patients entering the system, fewer doctors retiring and not being replaced at the same rate.

And crucially, each individual patient requires significantly more diagnostic work than they did like 20 years ago.

That last point is the real multiplier. I mean, decades ago, a pathologist might look at a slide under a microscope and make a relatively straightforward morphological assessment

like just saying yes this is cancer or no it's benign.

Exactly. Yes, this is invasive ductal carcinoma. But today oncology has become highly personalized. Oncologists don't just want to know if it's cancer.

Right. They need the specific molecular subtype,

right? So pathologists are constantly running and interpreting extra immunohistochemical stains plus molecular tests just to figure out exactly which targeted therapy will work.

So you have a shrinking pool of human experts doing um exponentially more complex cognitive work per patient.

Exactly. It's unsustainable.

So let's break down how technology is attempting to solve this. The first step the paper outlines is the transition from analog to digital.

Yes. The foundational step is digital pathology. This is the physical process of taking those traditional glass H& slides and running them through high throughput scanners

and that creates what we call whole slide images, right? Or WSI. Exactly. It scans them at incredibly high resolution.

Now, that sounds simple, but a single whole slide image is massive, isn't it? I mean, we're talking about gigapixels of data.

Oh, yeah. They are incredibly dense files. A single digital slide can actually take up more hard drive space than a featurelength film. It is wild.

But once that slide is converted into a matrix of digital pixels, you can feed that data into convolutional neural networks or CNN's.

Okay, wait. I want to pause on convolutional neural network. works for our trailblazers. How does a CNN actually see a gigapixel image of a tumor?

Well, it works by mimicking the human visual cortex, but at a microscopic mathematical level. A CNN doesn't look at the whole picture at once. It uses layers of mathematical filters. The first layer might scan the millions of pixels just looking for sharp contrasts and color, you know, edges,

right? Just basic outlines.

Exactly. Then the next layer takes those edges and looks for curves or basic shapes. Deeper layers combine those shapes to identify complex cellular structures like the nucleus of a cell.

Oh wow. So it's building up the image piece by piece,

right? It breaks the visual data down into thousands of mathematical features that it can measure simultaneously. Applying these algorithms gives birth to computational pathology.

You know, this brings to mind the transition from physical paper maps to modern GPS apps.

Oh, that's a good comparison.

Yeah, like digitizing a physical map into a PDF on your phone was the first step. That's digital pathology. You can view it on a screen, zoom in, whatever,

right?

But adding the algorithms that analyze millions of data points to predict traffic or reroute you around accidents, that analytical layer is the AI.

Yeah,

that is computational pathology.

That analogy captures the functional leap perfectly. And just like navigation apps help you avoid traffic jams, these AI algorithms are stepping in to assist pathologists in their daily trenches,

starting by automating the routine, tedious tasks. Right.

Exactly. The paper specifically highlights the Chamellon 16 trial regarding lymph nodes.

Right. Because when a patient has breast cancer, the surgeon often removes the sentininal lymph nodes first.

Yes. Because that's the first place the cancer is likely to spread. And pathologists have to slice these nodes up and meticulously search them for tiny clusters of metastatic cancer cells.

Which sounds like finding a needle in a haststack.

It's an incredibly demanding task. You're asking a human being to scan a massive microscopic landscape for a microscopic anomaly,

right? It's basically like examining a highdefinition satellite map of New York City and trying to find a single person wearing a red hat.

Oh my gosh, that sounds exhausting.

It is. And human eyes naturally get fatigued. But the Chameleon 16 trial showed that AI algorithms could detect these micrometastases with an accuracy that exceeded human pathologists.

Wait, really? The AI beat the humans?

Yes. Specifically, when those humans were placed under typical clinical time constraints in In real world applications, AI can cut the time a pathologist spends analyzing a single slide by up to 50%.

Okay, I have to ask a practical question here, though.

If the algorithm is incredibly sensitive, like designed to find absolutely every suspicious cluster, doesn't that risk overfling anomalies?

That's a very fair point

because if the computer is constantly tapping the doctor on the shoulder saying, "Look at this spot and this spot," it feels like it could actually create more work for the human who now has to verify every false alarm.

And that is a fundamental concern when designing any medical screening tool. You have to balance sensitivity with specificity,

right?

However, the data from these pathology algorithms shows they're engineered with an incredibly high negative predictive value or NPV.

Let's define negative predictive value for the listener real quick.

Sure. Negative predictive value is the probability that when the test says a sample is negative or healthy, it truly is negative. And in several studies, these AI models achieved an NPV approaching 100% for detecting things like lymph node metastasis.

Wow. So, it is exceptionally good at knowing when a slide is completely normal.

Yes. The real workflow benefit isn't necessarily finding the cancer faster. It's confidently clearing the cue of healthy tissue.

So, it's an ultra reliable screener.

Exactly. It essentially tells the pathologist, I have analyzed these 50 slides. They are completely clear. Do not waste your time here.

That is huge. That allows the human expert to focus their limited energy strictly on the complex ambiguous cases.

Precisely. And it also drastically reduces the need for hospitals to order expensive downstream chemical testing just to doublech checkck clearly negative cases.

That makes the clinical value really obvious. Yeah.

Now the paper also discusses AI assisting with mitosis counting which is part of the um and Ellis grading system.

Right. The ston and Ellis system is how we determine the hisytological grade of a breast tumor. Basically how aggressive it is. A crucial component of that grade is counting mitosis which are individual cells actively caught in the process of dividing.

So a pathologist has to look at a highly magnified area and literally tally up the dividing cells.

Yes. And catching a cell mid division just by its physical shape under a microscope is notoriously subjective.

I can imagine one person's dividing cell is another person's what? Smudge.

Exactly. Human reproducibility is low because one pathologist might interpret a distorted new nucleus as a dividing cell while another views it as just an artifact of the slide preparation.

So how does the AI help here?

It automates the process by highlighting the specific hotspots of cell division with mathematical consistency. The studies in the paper showed this AI assistance improved tumor grading accuracy by 15%.

Wow. And it probably massively increase the agreement between different pathologists reviewing the same case. Right.

Absolutely. It removes a lot of the desk work.

Okay. So the algorithm is great at saving time and counting visible shapes, but the paper outines is a second major hurdle that AI is tackling.

Yes, the limits of human visual perception when it comes to color,

right? What the paper calls the biomarker battle. Let's dig into that.

So, in breast cancer, diagnosing the disease is only half the battle. We rely heavily on iminohistochemistry or IHC scoring to determine the treatment plan.

And that involves bathing the tissue in antibodies. Right.

Correct. We bathe the tissue slice in specialized antibodies that stain specific proteins brown or red for Example, we look for Kai 67, which indicates cellular proliferation

or hormone receptors like estrogen and progesterone, the ER and PR.

Exactly. And the thing is, the human visual cortex is simply not designed to look at a field of thousands of cells and accurately calculate the exact percentage of them that have a faint brown tint.

Right? It's just a physiological limitation.

It is. There is notoriously high interobserver variability. One pathologist might estimate 15% of the cells are staining positive, while another specialist looking at the exact same slide might estimate 8%.

Which could totally change the treatment, I'm assuming.

Oh, absolutely. Especially when you're dealing with low expression ER or PR, say, trying to decide if less than 10% of cells are showing faint color. Human estimation becomes highly variable.

But the AI tools don't have that problem.

No, they achieve near perfect concordance here because they aren't estimating. They are quantifying the exact pixel values of the ain across the entire slide without fatigue or bias.

For you trailblazers listening, think of it like this. Relying on human vision for IHC scoring is like asking five people to eyeball a painted wall and name the exact shade of beige.

Yeah. You will get five different answers based on the room's lighting and their own perception.

Exactly. Using the AI is like pressing a perfectly calibrated digital colorimeter against that wall to instantly get the exact hex code. The Humei is estimating, but the machine is reading the math.

That colorimeter analogy is vital for understanding the current paradigm shift with the HR2 biioarker.

Ah Er2 let's talk about that.

So H2 or human epidermal growth factor receptor 2 is a protein that promotes cancer cell growth. Historically assessing it was a binary question

like either a strong positive or a negative.

Right. A tumor either had a massive amplification of this protein or it didn't. But the pharmaceutical landscape changed dramatically

because of the new drugs.

Exactly. We now have revolutionary drugs called anti-HR2 antibbody drug conjugates. They are essentially chemotherapy payloads attached to a homing beacon that seeks out the HR2 protein.

And the crazy thing is these drugs are incredibly effective even if the tumor has barely perceptible amounts of the HR2 protein on its surface. Right.

Exactly. And this created entirely new clinical categories called HR2 low and HR2 ultra low.

Wait, I want to anticipate what our listeners might be thinking here. Are we really saying we're going to trust a computer algorithm over a doctor with 20 years of experience to decide if a patient gets a targeted cancer drug?

It is a profound question, but it comes back to that physiological limit of the human eye. We are now asking human pathologists to distinguish between a score of zero, absolutely no stain, and a score of one plus,

which is faint, right?

Yeah, extremely faint. Barely perceptible staining in perhaps 10% of the cells. Distinguishing a 5% faint stain from a 0% stain across hundred of slides is something the human eye literally cannot do reliably.

So the risk of human misdiagnosis here is unacceptably high.

It is because missing that microscopic faint stain means denying a patient a life-saving therapy. And the machine just doesn't struggle with faint color.

What do the numbers say in the paper?

The data in the RUL paper shows AI reaching over 92% agreement in HR2 scoring, specifically excelling at distinguishing those faint zero from one plus cases.

It's incredible.

And we see the same leap in PDL1 scoring, which is a protein evaluated to determine if a patient with triple negative breast cancer should receive amunotherapy.

What was the human agreement on that?

Human concordance on PDL1 was sitting around a moderate 70%. When AI assists, that consistency jumps to 95%.

Wow. So, the AI acts as the perfect digital colorimeter for what is physically visible on the slide.

But, um, the truly mind-blowing section of this paper is what the AI can see that isn't visually there at all.

Yes. This is where computational pathology transcends traditional medicine. We are moving from diagnostic assistance into predictive modeling.

So using deep learning to predict treatment outcomes and internal genomic mutations straight from that standard cheap H& slide.

Exactly.

Just to be perfectly clear for the audience, we are not running new expensive chemical tests on the tissue here. The algorithm is just looking closely at the original pink and purple shapes and predicting DNA level data.

That is the reality. For instance, before surgery, Patients often receive neoagivant chemotherapy to shrink the tumor. If the tumor is completely eradicated by the drugs before surgery, it's called a pathological complete response or PCR. Achieving a PCR indicates a highly favorable long-term prognosis.

And the AI can predict this.

Yes, AI models can now analyze the basic pre-treatment biopsy slide and predict with high accuracy whether that specific patient's tumor will achieve a PCR.

It is predicting the future behavior of the tumor. just from the spatial patterns of the cells. That is wild.

Furthermore, the AI can infer expensive molecular alterations. Consider the BRCA1 and BRCA2 gene mutations which are critical for assessing hereditary cancer risk.

Normally, identifying those requires complex, costly genetic sequencing. Right.

Exactly. Because a mutation is an alteration deep inside the DNA of the cell. You shouldn't be able to see it on a cellular map.

But the AI does.

It does. It looks at the overall morphology, the architecture of the tissue and predicts BRCA mutations and something called homologous recombination deficiency or HRD.

Okay, let's define HRD for the audience because the mechanics of how the AI sees it are just fascinating.

So homologous recombination deficiency means the cancer cell has lost its ability to repair double strand breaks in its DNA.

Right. DNA breaks naturally and normal cells fix it.

Exactly. But when a tumor can't fix these breaks, its genome becomes chaotic. As the tumor grows with this broken repair mechanism, it subtly changes how it interacts. at the surrounding tissue.

So it leaves a macroscopic footprint of a microscopic error. What kind of footprint?

The AI recognizes specific architectural clues. For example, it identifies laminated fibrosis, which is a specific type of scar-like tissue laid down in distinct layers around the tumor.

Oh wow.

It also analyzes the density and pattern of tumor infiltrating lymphosytes. You know, the patients immune cells that are crowding around the chaotic cancer cells trying to attack them.

And the algorithm correlates these structural patterns to the broken DNA repair mechanism.

Yes, with astonishing accuracy.

This is staggering. It is like looking at a highresolution photograph of a baked cake's crumb structure and the AI correctly predicting the exact brand of flour that was used in the exact temperature of the oven.

That's a brilliant way to think about it.

It's extracting the invisible molecular recipe purely from the structural photograph.

The foundational biological principle here is that the genotype, the DNA recipe ultimately dictates the phenotype, which is of the physical structure.

So the changes were always there on the glass slide.

Exactly. It just required a convolutional neural network capable of analyzing millions of pixels simultaneously to notice the mathematical correlations.

Now the paper mentions this capability extends to complex genomic signatures as well. Specifically the enkotype DX test.

Yes. And this is considered a financial and clinical holy grail in oncology. Ankotype DX is a 21 gene assay used widely in early stage breast cancer. It generates a recurrence score from 0 to 100. Right.

Correct. This score tells oncologists whether the patient has a high risk of the cancer returning and crucially whether they actually need chemotherapy or if they can safely skip it.

But the anker type DX test costs thousands of dollars. It takes weeks to process and it physically consumes part of the tissue sample. Not every clinic globally can afford it or wait for the results.

That's the problem. But the paper highlights predictive models like Orpheus developed by a company called at AI.

Okay.

Orpheus is a deep learning tool designed to predict that exact encoype DX risk category directly from the digitized H& slide. In external validation tests, it achieved an area under the curve, an AU of89.

For our trailblazers, the area under the curve is a grading scale for predictive models. A score of.5 is basically no better than a coin toss, right?

A score of 1.0 is perfect prediction. So an AU of89 is essentially a solid agrade in accuracy.

It is excellent clinical accuracy. And astonishingly, in some retrospective studies, the AI algorithm actually outperformed the expensive molecular test itself in correctly identifying which patients eventually suffered a metastatic recurrence.

That is just incredible. But um if these tools can rapidly score biomarkers, predict drug responses, and magically infer genomic data from a $5 slide, we have to ground this excitement in reality for a second.

Yes, we do.

Why isn't this software running in every hospital in the world right now? What is the clinical friction keeping this in the journal club phase?

Well, we do have powerful prognostic tools developed right now. Algorithms like the brace marker, stratopath breast and rocks risk BC, right?

Robs risk BC, for instance, was validated on the massive Kanto trial cohort in France. Validating on a real world trial cohort proved that the algorithm could predict 5-year metastasis free survival better than traditional clinical models.

So, the data is there. What's the holdup?

The logistical barriers to everyday to use are steep. The primary hurdle is infrastructure cost.

Ah because even if the software saves you thousands of dollars on genetic tests later, you still need to digitize the slides today.

Exactly. A hospital must first purchase high throughput whole slide scanners which are incredibly expensive. They then have to upgrade their IT infrastructure to handle gigabit data transfers securely and purchase pabytes of server storage.

That initial capital investment is a massive roadblock, especially for midsize or smaller rural clinics

it is. Then there is the issue of scientific validation. Medicine relies on randomized clinical trials.

Right. You can't just change the standard of care based on a retrospective algorithm.

Exactly. You have to prove it perspectively. That means running the AI on new patients in real time and tracking their outcomes over years which is timeconuming and phenomenally expensive.

And the outline of the paper also points to the danger of AI bias right the need for diverse data sets to prevent health equity harms.

This is a critical technical limitation. If an algorithm is trained primarily on slides from a top tier research hospital in Paris, it learns the specific pink and purple hues generated by that hospital's specific chemical staining protocol.

Oh, and there's specific brand of slide scanner.

Exactly. So, if you deploy that same AI in a rural clinic with a different scanner and slightly different chemical dyes, the AI's accuracy can plummet. The training data must be globally diverse to be robust.

That makes total sense and that leads to the final and perhaps most difficult roadblock.

Yeah,

the blackbox problem.

Yes, the black box.

Because deep learning analyzes millions of parameters. It might output a score saying this patient has a 40% risk of recurrence, but a human pathologist cannot reverse engineer the neural network to see why it reached that conclusion.

It is a regulatory nightmare.

It raises a huge ethical dilemma for everyday practice. If you don't know the specific biological reason an algorithm is flagging a patient as high risk, how do you justify by adding a toxic chemotherapy drug to their treatment plan.

You can't.

Right? If a patient looks at you and asks, "Doctor, why am I getting this specific chemo?" You cannot just answer, "Well, the computer said so."

You cannot. And regulatory bodies simply will not allow it. To address this, the medical community recently released the Starred AI guidelines, the standards for reporting of diagnostic accuracy studies for AI.

What does those guidelines do?

It's a framework demanding developers transparently report their training data demographics, error rates, and clinical limitations. But the ultimate solution to the blackbox is moving toward interpretable agent-based AI.

What does agent-based AI look like in a clinical setting?

Imagine integrating advanced large language models with these visual algorithms. Instead of the AI just spitting out a static risk score on a PDF, it becomes an interactive consultant.

Okay, so during a multid-disciplinary tumor board, an oncologist could ask the software, why did you categorize this tumor as high? risk for recurrence.

Exactly. And the software responds in plain language.

Oh wow.

The AI could say something like, "Based on my analysis, I detected a high density of laminated fibrosis in quadrant 3 alongside an abnormal motic rate. In my training data on the Kanto trial, this specific combination of features strongly correlates with broken DNA repair mechanisms."

So it provides synthesized explainable evidence.

It does not dictate care. It defends its thesis so the human doctors can make an informed for irmed ethical decision.

That is just fascinating. Well, this has been a phenomenal look at the cutting edge of medicine.

It really is an exciting time for the field.

To summarize our journey for the trailblazers listening, we have seen how computational pathology is evolving in three distinct stages. First, it acts as a tireless screener clearing the busy work of healthy tissue. Right?

Second, it serves as an objective perfectly calibrated digital calimeter for complex subjective biomarkers like HR2. And finally, It is transforming into a predictive model extracting invisible genomic data right out of the cellular architecture.

It forces us to realize that human vision for all its evolutionary brilliance is no longer the ceiling for medical diagnostics. We are teaching algorithms to read the structural language of biology.

I love that. I want to thank all of you trailblazers for joining us on this deep dive into the source material. But before we sign off, I want to leave you with a final lingering question to ponder.

Yes.

As these algorithms become increasingly adept at predicting expensive complex genomic signatures straight from a basic cheap hematoxin and eosin slide will the multi,000 molecular sequencing assays of today eventually become obsolete

it's a huge question

or as we unlock the hidden depths of the standard biopsy will the very definition of what constitutes a basic slide fundamentally change the economics of global healthcare entirely

definitely something to think about

absolutely next time you look at a simple pink and purple stained slide ask yourself what What is the computer seeing that you aren't?